@warmdrift/kgauto-compiler 2.0.0-alpha.82 → 2.0.0-alpha.85

package/dist/index.mjs

@@ -19,7 +19,7 @@ import {
   LIBRARY_VERSION,
   createKeyHealthRoute,
   keyFingerprint
-} from "./chunk-L246YOV7.mjs";
+} from "./chunk-WXN7FNXP.mjs";
 import {
   ABSOLUTE_FLOOR,
   ARCHETYPE_FLOOR_DEFAULT,
@@ -27,6 +27,7 @@ import {
   COST_RANKING_REFERENCE_SHAPE,
   MEASURED_GROUNDING_MIN_N,
   PROVIDER_ENV_KEYS,
+  canonicalPolicySet,
   chainProviderSpread,
   configureBrainQuery,
   createBrainQueryCache,
@@ -52,9 +53,10 @@ import {
   loadArchetypePerfFromBrain,
   loadArchetypePerfNFromBrain,
   loadChainsFromBrain,
+  policySetHas,
   readBrainReadEnv,
   resolveProviderKey
-} from "./chunk-XPD4I3Q5.mjs";
+} from "./chunk-LO2JXTGG.mjs";
 import {
   ALIASES,
   LATENCY_TIER_MS,
@@ -64,8 +66,9 @@ import {
   getProfile,
   latencyTierOf,
   profilesByProvider,
+  resolveModelAlias,
   tryGetProfile
-} from "./chunk-VVRDFE6T.mjs";
+} from "./chunk-FD3NFXDC.mjs";
 import {
   emitAdvisoryFired,
   emitCompileDone,
@@ -767,8 +770,8 @@ function effectiveConventions(profile) {
 function passScoreTargets(ir, opts) {
   const constraints = ir.constraints ?? {};
   const policy = opts.policy ?? {};
-  const blockedSet = new Set(policy.blockedModels ?? []);
-  const preferredSet = new Set(policy.preferredModels ?? []);
+  const blockedSet = canonicalPolicySet(policy.blockedModels);
+  const preferredSet = canonicalPolicySet(policy.preferredModels);
   const scores = [];
   const policyMutations = [];
   const rawPromotion = opts.promotion;
@@ -791,7 +794,7 @@ function passScoreTargets(ir, opts) {
       continue;
     }
     const reasons = [];
-    if (blockedSet.has(modelId)) {
+    if (policySetHas(blockedSet, modelId)) {
       reasons.push(`blocked_by_policy (consumer gated this model \u2014 see CompilePolicy.blockedModels)`);
     }
     if (opts.estimatedInputTokens > profile.maxContextTokens * 0.9) {
@@ -821,7 +824,7 @@ function passScoreTargets(ir, opts) {
     const qualityScore = Math.max(0, baseQuality - qualityPenalty);
     const callerOrderBoost = (modelIds.length - modelIds.indexOf(modelId)) * 0.1;
     const costPenalty = estimatedCostUsd * 5;
-    const preferredBoost = preferredSet.has(modelId) ? 0.5 : 0;
+    const preferredBoost = policySetHas(preferredSet, modelId) ? 0.5 : 0;
     let latencyPenalty = 0;
     const maxLatencyMs = constraints.maxLatencyMs;
     if (typeof maxLatencyMs === "number" && maxLatencyMs > 0) {
@@ -878,7 +881,7 @@ function passScoreTargets(ir, opts) {
         description: `Model ${modelId} excluded \u2014 estimated cost $${estimatedCostUsd.toFixed(4)} exceeds policy ceiling $${policy.maxCostPerCallUsd.toFixed(4)}`
       });
     }
-    if (preferredSet.has(modelId) && reasons.length === 0) {
+    if (policySetHas(preferredSet, modelId) && reasons.length === 0) {
       policyMutations.push({
         id: `policy-preferred-${modelId}`,
         source: "compile_policy",
@@ -2361,6 +2364,67 @@ function advisorRuleCrossFamilyFit(ctx) {
   ];
 }
 
+// src/advisor-rules/blocked-model-drift.ts
+var BLOCKED_MODEL_NOT_IN_ROSTER_CODE = "blocked-model-not-in-roster";
+var BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE = "blocked-model-family-sibling-served";
+var DOCS_URL = "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#best-practice-advisories";
+function familyOf2(modelId, profile) {
+  return profile?.family ?? deriveFamilyFromModelId(modelId);
+}
+function isOnTheWayOut(status) {
+  return status === "legacy" || status === "deprecated";
+}
+function advisorRuleBlockedModelDrift(ctx) {
+  const blocked = ctx.policy?.blockedModels;
+  if (!blocked || blocked.length === 0) return [];
+  const resolve = ctx.resolveProfile ?? tryGetProfile;
+  const out = [];
+  const entries = [...new Set(blocked)].sort();
+  const selectedProfile = resolve(ctx.selectedModelId);
+  const selectedFamily = familyOf2(ctx.selectedModelId, selectedProfile);
+  const orphans = entries.filter((e) => resolve(e) === void 0);
+  if (orphans.length > 0) {
+    const list = orphans.map((o) => `\`${o}\``).join(", ");
+    const plural = orphans.length === 1 ? "entry" : "entries";
+    const verb = orphans.length === 1 ? "matches" : "match";
+    out.push({
+      level: "warn",
+      code: BLOCKED_MODEL_NOT_IN_ROSTER_CODE,
+      message: `CompilePolicy.blockedModels ${plural} ${list} ${verb} no model in the current roster, so ${orphans.length === 1 ? "it is" : "they are"} inert \u2014 \`blockedModels\` is matched by exact model id, and nothing kgauto can select carries ${orphans.length === 1 ? "that id" : "those ids"}. The block will never fire.`,
+      suggestion: `Check for a typo, or for an id that was retired from the roster since the block was written. This is how a block goes quiet without an error: tt-intel (2026-07-29) carried \`KGAUTO_BLOCKED_MODELS="claude-sonnet-4-6"\` across a roster retarget and the block stopped covering the traffic they believed it covered. Resolve the intended model id against the live roster \u2014 \`getRecommendedPrimary({ family, fallback })\` returns the id the family currently resolves to \u2014 and block that id, or drop the entry if it is no longer needed.`,
+      docsUrl: DOCS_URL
+    });
+  }
+  if (selectedFamily !== null) {
+    const siblings = entries.filter((e) => {
+      if (resolveModelAlias(e) === resolveModelAlias(ctx.selectedModelId)) return false;
+      const p = resolve(e);
+      if (familyOf2(e, p) !== selectedFamily) return false;
+      if (p && selectedProfile && p.provider !== selectedProfile.provider) {
+        return false;
+      }
+      return true;
+    });
+    if (siblings.length > 0) {
+      const list = siblings.map((s) => `\`${s}\``).join(", ");
+      const plural = siblings.length === 1 ? "" : "s";
+      const retargetShaped = siblings.some((s) => {
+        const p = resolve(s);
+        return isOnTheWayOut(p?.status) && selectedProfile?.status === "current";
+      });
+      const retargetNote = retargetShaped ? ` The blocked id${plural} ${siblings.length === 1 ? "is" : "are"} legacy/deprecated while \`${ctx.selectedModelId}\` is current \u2014 that is the signature of a roster lifecycle move rather than a deliberate per-generation block, so this is more likely to be drift than intent.` : "";
+      out.push({
+        level: "warn",
+        code: BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE,
+        message: `\`${ctx.selectedModelId}\` was selected for this call. It is in the same family (\`${selectedFamily}\`) as blocked entr${siblings.length === 1 ? "y" : "ies"} ${list}, but its exact id differs \u2014 and \`CompilePolicy.blockedModels\` matches by exact id, so the block does not cover it.${retargetNote}`,
+        suggestion: `Two readings, and kgauto cannot tell them apart: (1) intentional \u2014 you meant to gate that specific id and \`${ctx.selectedModelId}\` is fine, in which case nothing needs doing and you can filter this code; (2) drift \u2014 you meant to gate the family, and a roster change moved traffic to a sibling your block never named. This happened to tt-intel: when the \`claude-sonnet\` family primary retargeted from \`claude-sonnet-4-6\` to \`claude-sonnet-5\`, their literal-id gate stopped matching the family's routed traffic (their fix was local family resolution). If you meant the family, add \`${ctx.selectedModelId}\` to \`blockedModels\` \u2014 kgauto deliberately does NOT widen exact-id blocks into family globs, because that would silently change what every existing block covers.`,
+        docsUrl: DOCS_URL
+      });
+    }
+  }
+  return out;
+}
+
 // src/advisor.ts
 var QUALITY_FLOOR_FOR_RECOMMENDATION = 6;
 var TIER_DOWN_COST_RATIO = 0.5;
@@ -2379,6 +2443,12 @@ function runAdvisor(ir, result, profile, policy, phase2) {
   out.push(...detectToolBloat(ir, result));
   out.push(...detectHistoryUncached(ir, profile));
   out.push(...detectSingleModelArray(ir, policy));
+  out.push(
+    ...advisorRuleBlockedModelDrift({
+      policy,
+      selectedModelId: profile.id
+    })
+  );
   if (policy?.posture !== "locked") {
     out.push(...detectCostMismatchedArchetype(ir, profile, phase2));
     out.push(...detectModelStaleEvidence(ir, profile));
@@ -2490,18 +2560,27 @@ function detectHistoryUncached(ir, profile) {
 function detectSingleModelArray(ir, policy) {
   if (ir.models.length !== 1) return [];
   if (policy?.posture === "locked") return [];
-  const only = ir.models[0];
+  const entry = ir.models[0];
+  const only = typeof entry === "string" ? entry : `family:${entry.family}`;
+  const blocked = canonicalPolicySet(policy?.blockedModels);
+  let alternatives = [];
+  try {
+    alternatives = getDefaultFallbackChain({
+      archetype: ir.intent.archetype,
+      primary: only,
+      posture: "preferred",
+      policy
+    }).filter((id) => resolveModelAlias(id) !== resolveModelAlias(only)).filter((id) => !policySetHas(blocked, id)).filter((id) => getModelCompatibility(id, { archetype: ir.intent.archetype }).status !== "reject");
+  } catch {
+  }
+  const hasAlternative = alternatives.length > 0;
+  const remedy = "Widen the chain AND pin your primary: `compile({ ...ir, models: getDefaultFallbackChain({ archetype: ir.intent.archetype, primary: '" + only + "', posture: 'preferred' }) }, { policy: { preferredModels: ['" + only + "'] } })`. The chain is a CANDIDATE SET \u2014 compile() scores it and does not honour its order, so without the `preferredModels` pin the extra entries can retarget your primary (and with it your cost profile). With the pin, `" + only + "` stays primary and the added entries serve only as the safety net. Note `posture` has no effect once `primary` is passed. If you gate models by spend, re-verify `policy.blockedModels` against the widened set before shipping. If single-model is intentional (compliance/brand promise), set `policy.posture = 'locked'` to silence this rule.";
   return [
     {
-      level: "warn",
+      level: hasAlternative ? "critical" : "warn",
       code: "single-model-array",
-      message: `\`ir.models\` has length 1 (only "${only}") and posture is not 'locked'. A single-model chain has no safety net \u2014 the first 429 / 5xx / cliff hits the user as a failure. Master plan \xA71.2 closes the reliability gap with a 2-step minimum.`,
-      // alpha.80: `posture: 'preferred'` keeps the consumer's current model at
-      // position 0 and is therefore cost-neutral; `'open'` re-picks the
-      // primary and is NOT. This rule is about reliability, not cost, so both
-      // stay on offer — but the cost consequence of the second is now stated,
-      // since a reliability fix should not silently become a repricing.
-      suggestion: "Use `getDefaultFallbackChain({ archetype: ir.intent.archetype, primary: '" + only + "', posture: 'preferred' })` for a user-anchored chain \u2014 this keeps `" + only + "` as your primary and only adds fallbacks, so your cost profile is unchanged. `getDefaultFallbackChain({ archetype, posture: 'open' })` instead lets the library pick the PRIMARY, ordered by archetype performance rather than cost \u2014 check what it returns before adopting it. If single-model is intentional (compliance/brand promise), set `policy.posture = 'locked'` to silence this rule.",
+      message: hasAlternative ? `\`ir.models\` has length 1 (only "${only}") for archetype "${ir.intent.archetype}" and posture is not 'locked'. A single-model chain has no safety net \u2014 the first 429 / 5xx / cliff hits the user as a failure \u2014 and a compatible, non-blocked alternative exists in the roster today (${alternatives[0]}), so this is critical: the missing safety net is adoptable now.` : `\`ir.models\` has length 1 (only "${only}") for archetype "${ir.intent.archetype}" and posture is not 'locked'. A single-model chain has no safety net \u2014 the first 429 / 5xx / cliff hits the user as a failure. No compatible non-blocked alternative is visible in the roster for this archetype, so this stays a warning.`,
+      suggestion: remedy,
       docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#single-model-array"
     }
   ];
@@ -3055,6 +3134,15 @@ function compile(ir, opts = {}) {
       sectionRewritesApplied
     }
   );
+  if (ir["policy"] !== void 0 && opts.policy === void 0) {
+    rawAdvisories.push({
+      level: "critical",
+      code: "policy-in-ir-ignored",
+      message: "The IR passed to compile() carries a `policy` field. `policy` belongs in the SECOND argument \u2014 `compile(ir, { policy })` \u2014 and inside the IR it is an unknown field that is completely ignored. If that policy names blockedModels, no block is being enforced on this call.",
+      suggestion: "Move it: `compile(ir, { policy: { ... } })`. If you also configure policy correctly elsewhere on this path, remove the IR copy so the next reader is not misled.",
+      docsUrl: "https://github.com/stue/command-center/blob/main/interfaces/kgauto.md#public-api"
+    });
+  }
   const advisories = rawAdvisories.map((a) => ({
     ...a,
     kgautoRequestId: handle,
@@ -4824,11 +4912,11 @@ async function call(ir, opts = {}) {
   }
   let policyBlockedFiltered;
   if (opts.policy?.blockedModels && opts.policy.blockedModels.length > 0) {
-    const blocked = new Set(opts.policy.blockedModels);
+    const blocked = canonicalPolicySet(opts.policy.blockedModels);
     const filtered = [];
     const dropped = [];
     for (const t of targetsToTry) {
-      if (blocked.has(t)) {
+      if (policySetHas(blocked, t)) {
         dropped.push(t);
       } else {
         filtered.push(t);
@@ -4868,10 +4956,18 @@ async function call(ir, opts = {}) {
   const failedProviders = /* @__PURE__ */ new Map();
   const sameModelRetryEnabled = opts.sameModelRetry ?? isSameModelRetryEnabledFromEnv();
   let retriedSameModel = false;
+  const pushAttempt = (attempt) => {
+    attempts.push(attempt);
+    if (attempt.status === "success") return;
+    try {
+      opts.onFailedAttempt?.(attempt);
+    } catch {
+    }
+  };
   for (let i = 0; i < targetsToTry.length; i++) {
     const targetModel = targetsToTry[i];
     if (opts.abortSignal?.aborted) {
-      attempts.push({
+      pushAttempt({
         model: targetModel,
         status: "terminal",
         errorCode: "aborted",
@@ -4882,7 +4978,7 @@ async function call(ir, opts = {}) {
     const targetProfile = tryGetProfile(targetModel);
     const providerFailReason = targetProfile ? failedProviders.get(targetProfile.provider) : void 0;
     if (targetProfile && providerFailReason && !opts.noFallback) {
-      attempts.push({
+      pushAttempt({
         model: targetModel,
         status: "terminal",
         errorCode: `${providerFailReason}_inferred`,
@@ -4901,7 +4997,7 @@ async function call(ir, opts = {}) {
           opts
         );
       } catch (err) {
-        attempts.push({
+        pushAttempt({
           model: targetModel,
           status: "terminal",
           errorCode: "compile_error",
@@ -4939,7 +5035,7 @@ async function call(ir, opts = {}) {
     }
     if (!validated.ok && isStructuredContractViolation(validated.errorCode) && sameModelRetryEnabled && !retriedSameModel && !retrySuppressionNote) {
       retriedSameModel = true;
-      attempts.push({
+      pushAttempt({
         model: targetModel,
         status: validated.errorType,
         errorCode: validated.errorCode,
@@ -4954,7 +5050,7 @@ async function call(ir, opts = {}) {
       servedByRetry = true;
     }
     if (validated.ok) {
-      attempts.push({
+      pushAttempt({
         model: targetModel,
         status: "success",
         ...servedByRetry ? { sameModelRetry: true } : {}
@@ -5072,7 +5168,7 @@ async function call(ir, opts = {}) {
         advisories: activeCompile.advisories
       };
     }
-    attempts.push({
+    pushAttempt({
       model: targetModel,
       status: validated.errorType,
       errorCode: validated.errorCode,
@@ -5652,6 +5748,13 @@ function combineOrderSwappedVerdicts(run1, run2) {
   if (run1 === run2) return run1;
   return "tied";
 }
+var ALT_STRATEGY_IDS = [
+  "discipline-gates-v1-alt",
+  "discipline-gates-v1-alt-blind"
+];
+function isAltStrategy(id) {
+  return id !== void 0 && ALT_STRATEGY_IDS.includes(id);
+}
 function replayRetryDelayMs(errorCode) {
   return errorCode === "rate_limit" ? 65e3 : 2e3;
 }
@@ -5705,7 +5808,34 @@ function altGatesBlockFor(args) {
     ...shapeAltering
   ]);
 }
+var ALT_BLIND_TOKEN_BUDGET_BREACH = {
+  ceilingTokens: 264,
+  measuredTokensTextWithTools: 285,
+  note: "Blind arm exceeds the stated 264-token ceiling on text+tools (285) and costs ~2.4x the v1-alt arm. Not corrected: compressing it would require editing prose the arm depends on NOT having been edited by this seat. Read a blind-arm LOSS as confounded by token tax; a WIN is unaffected."
+};
+var DISCIPLINE_GATES_V1_ALT_BLIND_HEADER = "Before stating any conclusion, run this check:";
+var ALT_BLIND_BULLETS = {
+  falsify: "- Before committing to an approach, state in one line what evidence would prove it wrong and how you would notice it. If nothing could falsify the approach, you hold a preference, not a plan \u2014 rework it until something could.",
+  deviation: "- When you choose not to follow an applicable instruction, say so explicitly: name the instruction, why you are deviating, and what you are doing instead. Silent deviation is forbidden \u2014 an override is only legitimate when it is visible.",
+  toolPredict: "- When using tools: before each call, state the result you expect. On mismatch, treat the gap as evidence your model of the system is wrong \u2014 revise the model before acting again; never silently retry.",
+  verifyDelegated: "- Verify every delegated or sub-agent result before composing it into an answer: spot-check it against the source or an independent probe. An unverified sub-result is a claim you are repeating, not a fact you know.",
+  markAssumed: '- In the final answer, keep verified claims and assumptions visibly distinct: mark anything unchecked as "assumed" or "unverified". Never let an assumption borrow the confidence of the verified claims beside it.'
+};
+function altBlindGatesBlockFor(args) {
+  const parts = [
+    "falsify",
+    "deviation",
+    ...args.hasTools ? ["toolPredict"] : [],
+    "verifyDelegated",
+    ...args.outputMode === "text" ? ["markAssumed"] : []
+  ];
+  return [
+    DISCIPLINE_GATES_V1_ALT_BLIND_HEADER,
+    ...parts.map((p) => ALT_BLIND_BULLETS[p])
+  ].join("\n");
+}
 var STRATEGY_ALT_SECTION_ID = "__kgauto_strategy_eval_gates_alt__";
+var STRATEGY_ALT_BLIND_SECTION_ID = "__kgauto_strategy_eval_gates_alt_blind__";
 function withAltDisciplineContract(ir) {
   const outputMode = resolveOutputMode({
     declared: ir.constraints?.outputMode,
@@ -5724,6 +5854,28 @@ function withAltDisciplineContract(ir) {
     ]
   };
 }
+function withAltBlindDisciplineContract(ir) {
+  const outputMode = resolveOutputMode({
+    declared: ir.constraints?.outputMode,
+    structuredOutput: ir.constraints?.structuredOutput,
+    toolCount: ir.tools?.length ?? 0
+  });
+  const hasTools = (ir.tools?.length ?? 0) > 0;
+  return {
+    ...ir,
+    sections: [
+      ...ir.sections ?? [],
+      {
+        id: STRATEGY_ALT_BLIND_SECTION_ID,
+        text: altBlindGatesBlockFor({ outputMode, hasTools })
+      }
+    ]
+  };
+}
+var ALT_SECTION_ID_BY_STRATEGY = {
+  "discipline-gates-v1-alt": STRATEGY_ALT_SECTION_ID,
+  "discipline-gates-v1-alt-blind": STRATEGY_ALT_BLIND_SECTION_ID
+};
 async function runGoldenEval(opts) {
   const fetchFn = opts.fetchImpl ?? fetch;
   const progress = opts.onProgress ?? (() => {
@@ -5791,7 +5943,9 @@ async function runGoldenEval(opts) {
   }
   const armBIr = (ir) => {
     if (axis !== "strategy") return ir;
-    return strategyId === "discipline-gates-v1-alt" ? withAltDisciplineContract(ir) : withDisciplineContract(ir);
+    if (strategyId === "discipline-gates-v1-alt") return withAltDisciplineContract(ir);
+    if (strategyId === "discipline-gates-v1-alt-blind") return withAltBlindDisciplineContract(ir);
+    return withDisciplineContract(ir);
   };
   const replay = async (ir, model, captureGates = false) => {
     let gates;
@@ -5803,13 +5957,14 @@ async function runGoldenEval(opts) {
       };
       const compiled = compile(evalIr);
       if (captureGates) {
-        if (strategyId === "discipline-gates-v1-alt") {
+        if (isAltStrategy(strategyId)) {
+          const altSectionId = ALT_SECTION_ID_BY_STRATEGY[strategyId];
           const altSection = (evalIr.sections ?? []).find(
-            (s) => s.id === STRATEGY_ALT_SECTION_ID
+            (s) => s.id === altSectionId
           );
           gates = {
             fired: altSection !== void 0 && altSection.text.length > 0,
-            rule: "discipline-gates-v1-alt",
+            rule: strategyId,
             gateTokens: altSection ? countTokens(`${altSection.text}
 
 `) : 0
@@ -6240,7 +6395,8 @@ function classifyStrategyOutcome(r) {
   if (r.losses > r.wins) return "loses";
   return "ties";
 }
-var STRATEGY_AUTHORSHIP_LIMITATION = "Both gate wordings share one author (kgauto-Cairn); a shared blind spot is invisible to this experiment. Strong verdicts: both-lose (mechanism) and v1-loses/alt-wins (wording). An independently-authored alt would strengthen every other cell.";
+var STRATEGY_AUTHORSHIP_LIMITATION = "Only the v1 wording was measured on this run; no re-wording arm was triggered, so this result says nothing about whether the wording or the mechanism produced it.";
+var STRATEGY_AUTHORSHIP_INDEPENDENT = "Three arms: v1 and one re-wording by this seat, plus an independently authored arm (context-free subagent, no repo access, no knowledge of the experiment). A both-lose verdict therefore spans independent authorship and is not explained by one author's blind spot. Residual limitation: all three arms encode a similar set of underlying principles, so a wrong CHOICE of principles remains harder to see than a wrong wording of them.";
 async function runStrategyEvalWithAttribution(opts) {
   const primary = await runGoldenEval({
     ...opts,
@@ -6257,9 +6413,20 @@ async function runStrategyEvalWithAttribution(opts) {
     axis: "strategy",
     strategy: "discipline-gates-v1-alt"
   });
-  const altOutcome = classifyStrategyOutcome(alt);
-  const attribution = altOutcome === "wins" ? "wording-failure" : altOutcome === "loses" ? "mechanism-failure" : "wording-inconclusive";
-  return { attribution, primary, alt, limitation: STRATEGY_AUTHORSHIP_LIMITATION };
+  const altBlind = await runGoldenEval({
+    ...opts,
+    axis: "strategy",
+    strategy: "discipline-gates-v1-alt-blind"
+  });
+  const outcomes = [classifyStrategyOutcome(alt), classifyStrategyOutcome(altBlind)];
+  const attribution = outcomes.includes("wins") ? "wording-failure" : outcomes.every((o) => o === "loses") ? "mechanism-failure" : "wording-inconclusive";
+  return {
+    attribution,
+    primary,
+    alt,
+    altBlind,
+    limitation: STRATEGY_AUTHORSHIP_INDEPENDENT
+  };
 }
 
 // src/oracle.ts
@@ -6516,6 +6683,30 @@ function createDelegate(opts) {
 }
 
 // src/advisories-api.ts
+var BURST_SPAN_MS = 5 * 60 * 1e3;
+function classifyEvidenceWindow(row) {
+  const { evidence_first_at: first, evidence_last_at: last, evidence_n: n } = row;
+  if (typeof first !== "string" || typeof last !== "string") return null;
+  if (typeof n !== "number" || !Number.isFinite(n) || n <= 0) return null;
+  const t0 = Date.parse(first);
+  const t1 = Date.parse(last);
+  if (Number.isNaN(t0) || Number.isNaN(t1)) return null;
+  const spanMs = Math.max(0, t1 - t0);
+  return { firstAt: first, lastAt: last, n, spanMs, isBurst: spanMs < BURST_SPAN_MS };
+}
+function formatEvidenceSpan(spanMs) {
+  if (spanMs < 1e3) return `${spanMs}ms`;
+  const s = Math.round(spanMs / 1e3);
+  if (s < 90) return `${s}s`;
+  const m = Math.round(s / 60);
+  if (m < 90) return `${m}m`;
+  const h = Math.round(m / 60);
+  if (h < 48) return `${h}h`;
+  return `${Math.round(h / 24)}d`;
+}
+function burstCaveat(w) {
+  return `Evidence is ${w.n} observation${w.n === 1 ? "" : "s"} spanning ${formatEvidenceSpan(w.spanMs)} (${w.firstAt} \u2192 ${w.lastAt}) \u2014 that is a single burst, not a standing rate. Check it against your own deploy log before acting: a cluster this tight is usually one incident, and may already be fixed.`;
+}
 var SEVERITY_SET = /* @__PURE__ */ new Set(["info", "warn", "critical"]);
 var STATUS_SET = /* @__PURE__ */ new Set(["open", "snoozed", "resolved"]);
 var RESOLUTION_SOURCE_SET = /* @__PURE__ */ new Set([
@@ -6555,10 +6746,18 @@ function rowToAdvisory(row) {
     if (docsLink) suggestedFix.docsLink = docsLink;
     if (suggestion) suggestedFix.before = suggestion;
   }
+  const evidenceWindow = classifyEvidenceWindow(
+    row
+  );
+  const declaredSeverity = asSeverity(row.severity);
+  const downgraded = evidenceWindow?.isBurst === true && declaredSeverity === "critical";
+  const effectiveSeverity = downgraded ? "warn" : declaredSeverity;
   const out = {
     id: typeof row.id === "string" ? row.id : "",
     rule: typeof row.rule === "string" ? row.rule : "",
-    severity: asSeverity(row.severity),
+    severity: effectiveSeverity,
+    evidenceWindow,
+    ...downgraded ? { severityBeforeBurstDowngrade: declaredSeverity } : {},
     openedAt: typeof row.opened_at === "string" ? row.opened_at : "",
     lastObservedAt: typeof row.last_observed_at === "string" ? row.last_observed_at : "",
     observationCount: typeof row.observation_count === "number" ? row.observation_count : 0,
@@ -6566,7 +6765,13 @@ function rowToAdvisory(row) {
       ...archetype ? { archetype } : {},
       ...model ? { model } : {}
     },
-    message: typeof row.message === "string" ? row.message : "",
+    // The caveat goes INLINE in the message, not only in `evidenceWindow`.
+    // A structured field a consumer must think to read is the same defect
+    // one layer along: the substrate knowing and not telling. The whole
+    // point is that the check costs a glance rather than a query.
+    message: (typeof row.message === "string" ? row.message : "") + (evidenceWindow?.isBurst ? `
+
+${burstCaveat(evidenceWindow)}` : ""),
     suggestedFix,
     autoApplicable: false,
     // reserved — alpha.30+
@@ -7103,9 +7308,14 @@ export {
   ABSOLUTE_FLOOR,
   ALIASES,
   ALL_ARCHETYPES,
+  ALT_BLIND_TOKEN_BUDGET_BREACH,
+  ALT_STRATEGY_IDS,
   ARCHETYPE_FAMILY_FITS,
   ARCHETYPE_FLOOR_DEFAULT,
+  BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE,
+  BLOCKED_MODEL_NOT_IN_ROSTER_CODE,
   BRAIN_READ_ENV_NAMES,
+  BURST_SPAN_MS,
   COACH_CFG,
   COST_RANKING_REFERENCE_SHAPE,
   CallError,
@@ -7116,6 +7326,7 @@ export {
   DEFAULT_PROMOTIONS_ENDPOINT,
   DELEGATE_TOOL_DEFINITION,
   DIALECT_VERSION,
+  DISCIPLINE_GATES_V1_ALT_BLIND_HEADER,
   DISCIPLINE_GATES_V1_ALT_HEADER,
   FamilyResolutionError,
   INTENT_ARCHETYPES,
@@ -7132,6 +7343,7 @@ export {
   RULE_DISCIPLINE_GATES_V1,
   RULE_DISCIPLINE_GATES_V1_STRUCTURED,
   RULE_SEQUENTIAL_TOOL_CLIFF,
+  STRATEGY_AUTHORSHIP_INDEPENDENT,
   STRATEGY_AUTHORSHIP_LIMITATION,
   TRANSLATOR_FLOOR,
   _testResetMeasuredFailure,
@@ -7139,6 +7351,7 @@ export {
   _testWaitForMeasuredFailureRefresh,
   _testWaitForPromotionsRefresh,
   allProfiles,
+  altBlindGatesBlockFor,
   altGatesBlockFor,
   applyArchetypeConvention,
   applySectionRewrites,
@@ -7152,9 +7365,11 @@ export {
   buildLLMJudge,
   buildPairwiseJudgePrompt,
   buildShadowProbeRow,
+  burstCaveat,
   call,
   captureGoldenIr,
   chainProviderSpread,
+  classifyEvidenceWindow,
   classifyStrategyOutcome,
   clearBrain,
   combineOrderSwappedVerdicts,
@@ -7174,6 +7389,7 @@ export {
   execute,
   findBetterFit,
   flushBrainDeadLetter,
+  formatEvidenceSpan,
   getActionableAdvisories,
   getAllStarterChains,
   getAllStarterChainsWithGrounding,
@@ -7195,6 +7411,7 @@ export {
   getStarterChainWithGrounding,
   hasMutation,
   hashShape,
+  isAltStrategy,
   isArchetype,
   isAutoPromoteEnabledFromEnv,
   isBrainQueryActiveFor,
@@ -7238,8 +7455,10 @@ export {
   resetTokenizer,
   resolveConventionsForProfile,
   resolveGoldenCaptureRate,
+  resolveModelAlias,
   resolvePricingAt,
   resolveProviderKey,
+  rowToAdvisory,
   rubricFor,
   runAdvisor,
   runGoldenEval,
@@ -7248,6 +7467,7 @@ export {
   shouldCaptureGolden,
   tryGetProfile,
   wilsonLowerBound,
+  withAltBlindDisciplineContract,
   withAltDisciplineContract,
   withDisciplineContract
 };

package/dist/{ir-Cx9hJj0B.d.ts → ir-DfrM71pA.d.ts}

@@ -305,6 +305,17 @@ interface CompilePolicy {
      * Model IDs the consumer has gated. Compile() will never select these.
      * Use for: cost caps, account-level rate limits, "this model is broken
      * for our workload" decisions.
+     *
+     * Matching is by EXACT id after alias resolution (alpha.85): an alias
+     * and its canonical id are the same model, so blocking either blocks
+     * both (`deepseek-chat` ≡ `deepseek-v4-flash`). Matching does NOT widen
+     * to family siblings — blocking `deepseek-v4-flash` says nothing about
+     * `deepseek-v4-pro` — and a bare family key (`'deepseek'`,
+     * `'claude-sonnet'`) is INERT: it matches no model and no warning fires
+     * today. To gate a family, list its concrete ids. Beware roster
+     * retargets: when a family primary moves to a new id, a literal-id
+     * block does not follow it (tt-intel s113 built consumer-side family
+     * resolution for exactly this).
      */
     blockedModels?: string[];
     /**
@@ -1073,6 +1084,44 @@ interface ShadowProbeConfig {
     skipSlowTierInSync?: boolean;
 }
 interface CallOptions {
+    /**
+     * alpha.84 — fires once per FAILED attempt during the fallback walk, before
+     * the walk continues. IC-Cairn's filing, 2026-07-29.
+     *
+     * ## The gap this closes
+     *
+     * `call()` records exactly ONE outcome row per call. An attempt that fails
+     * and then walks to a successful fallback survives only as
+     * `fellOverFrom`/`fallbackReason` metadata ON the success row — so a
+     * `call()`-only consumer's corpus contains **no failure rows at all**, by
+     * construction, no matter how carefully they wired `record()`.
+     *
+     * That is not hypothetical. kgauto's own `corpus-implausible-success`
+     * detector fired on inspire-central and told them to "wire record() on the
+     * FAILURE path" — which they had already done, in five places. Measured:
+     * 89 rows, 0 failure rows, 2 fellover-bearing. **The detector demanded
+     * output the API could not produce.** Streaming consumers have had
+     * `onFailedAttempt` since alpha.48 via `streamWithFallover`; `call()`
+     * consumers had nothing. This restores parity.
+     *
+     * ## Contract
+     *
+     * Fires for every attempt whose `status` is `'retryable'` or `'terminal'`,
+     * including skipped ones (`*_inferred`, `aborted`) — a skip is a real
+     * datum about the chain, not an absence. Never fires for `'success'`.
+     *
+     * Called synchronously and **its throw is swallowed**: a consumer's
+     * telemetry bug must never take down the call it is observing. It is a
+     * notification, not a hook that can veto or alter the walk.
+     *
+     * Typical use — emit a failure row per attempt so the corpus can contain
+     * failures at all:
+     *
+     *   call(ir, { onFailedAttempt: (a) => void recordOutcome({
+     *     ...base, success: false, errorType: a.errorCode, model: a.model,
+     *   }) })
+     */
+    onFailedAttempt?: (attempt: CallAttempt) => void;
     /** Forwarded to compile(). */
     policy?: CompilePolicy;
     /**